Intelligent Language Processing for Understanding Financial Text

该创新奖学金旨在解决智能语言处理的主题，以支持特定业务合作伙伴的具体需求，并为英国工业产生更广泛的积极影响。该奖学金面向语言问题，作为机器人和人工智能（AI）领域的“难题”，这些是英国工业战略确定的技术挑战。“难题”是，对语言的理解不仅植根于文本或演讲本身，而且植根于语境——包括人类沟通者拥有的世界知识，但在很大程度上缺乏人工智能。因此，从人工智能的角度来看，语言理解是自动化系统最具挑战性的更高层次的认知任务。以“银行”这个词为例。要在文本（书面或口头）中正确处理这一词的使用，首先有必要确定其语法功能。这是相对简单的(“I bank at HSBC”-动词；“汇丰是一家全球性银行”（名词）。区别含义就不那么简单了——例如，金融“银行”和河流“银行”。对于人工智能来说，更困难的是区分相同语法形式和相同基本含义的情况，因为含义/含义的微妙性而必须以不同的方式解释：考虑“我把钱存在银行”（作为客户）/“我在银行赚了钱”（作为员工）/“我在银行丢了钱”（银行是一个物理位置，而不是一个公司实体）。多个研究领域已经开发出在某种程度上解决这一问题的方法，使计算机辅助语言理解成为可能。这些技术被称为基于语料库的方法，或称为自然语言处理。它们涉及应用数字技术：(i)通过基于非常大的语言数据集的机器学习来训练计算机完成分析任务；或者（ii）对如此大的数据集进行抽样和筛选，以指导人类分析师找到他们所寻求的细节。这些都是智能语言处理的变种：也就是说，使用语言的详细知识，通过机器驱动的文本“大数据”分析，来驱动增强的基于语言的人工智能，以进一步实际利用这些数据。该奖学金解决的关键挑战是需要使这些尖端的人工智能技术为商业和工业带来变革，不仅为整体工业战略提供服务，还为我们的主要私营部门合作伙伴毕马威（KPMG）和监管部门的其他二级合作伙伴的业务需求提供服务。总的来说，该方法是将智能语言处理应用于理解公司财务环境中产生的文本类型的任务-财务文本。为了支持商业影响和知识交流，将解决两个主要问题：首先-从公司制作的文本文件中提取数字信息。（以实际英镑计算，一家公司发表的叙述能告诉我们他们的财务状况吗？）第二，研究如何在不同渠道间比较财务文本。（公司如何在针对不同受众的文件中以不同的方式叙述自己的地位，这又能告诉我们他们的实际地位是什么？）这些问题将通过本奖学金的三个研究工作项目来解决。(1)针对不同的受众将不同的文本联系在一起，理解语气和积极/消极。(2)放眼国际——在不同的法律/文化背景下，公司文件在世界各地有何不同？(3)信任/影响——公司如何在与不同受众的沟通中使用语言来说服和建立信任？

This Innovation Fellowship addresses the theme of Intelligent Language Processing in support of concrete needs for a specific business partner - as well as generating positive impact for UK industry more broadly. The fellowship is oriented to the issue of language as a 'hard problem' within the sphere of robotics and artificial intelligence (AI), these being a technological challenge identified by the UK Industrial Strategy. The 'hard problem' is that the understanding of language is rooted not merely in the words of the text or speech itself but also in the context - including world knowledge possessed by human communicators but largely lacking in AI. Language understanding is thus, from the perspective of AI, a higher level cognitive task of the kind most challenging for automated systems. Consider, for instance, the word ''bank''. To properly handle the use of such a word in a text (written or spoken), it is first necessary to identify its grammatical function. This is relatively straightforward (''I bank at HSBC'' - verb; vs. ''HSBC is a global bank'' - noun). Less straightforward is distinguishing meanings - a financial ''bank'' versus a river ''bank'', for instance. Even more difficult for AI is distinguishing cases where the same grammatical form and same basic meaning must be interpreted differently due to subtleties of meaning/implication: consider ''I put my money in the bank'' (as a customer) / ''I made my money in the bank'' (as an employee) / ''I lost my money in the bank'' (a bank as a physical location rather than a corporate entity).Multiple research fields have developed methods that address this issue to some extent, allowing computer-assisted language understanding. These techniques are variously referred to corpus-based methods, or as natural language processing. They involve applying digital technology either (i) to train computers to accomplish analytic tasks via machine learning based on very large language datasets; or (ii) to down-sample and sift such large datasets to guide a human analyst to the details they seek. These are all varieties of Intelligent Language Processing: that is, using detailed knowledge of language, derived via machine-driven analysis of textual 'big data', to drive enhanced language-based AI for practical exploitation of yet further such data. The key challenge this Fellowship addresses is the need to make these cutting-edge AI techniques transformative for business and industry, delivering not only for the overall Industrial Strategy, but also for the business needs of our primary private sector partner: KPMG and additional secondary partners in the regulatory sector. The method overall is to apply Intelligent Language Processing to the task of understanding types of text produced in corporate financial contexts - Financial Text. TWO major problems will be addressed in support of business impact and knowledge exchange: First - the extraction of numerical information from textual documents produced by companies. (What can the narratives a firm publishes tell us about their financial situation, in actual pounds)? Second - working out how to compare financial text between different channels. (How do companies narrate their status differently in documents directed to different audiences, and what does that tell us about their actual status? )These problems will be addressed through THREE research work projects in this Fellowship. (1) Linking together different texts for different audiences, and understanding tone and positivity/negativity. (2) Looking internationally - how do corporate documents vary around the world, in different legal / cultural context. (3) Trust / influence - how do companies use language to presuade and establish trust in their communications to different audiences?